Thought Machine and Amazon Web Services are applying AI agents to one of banking technology’s more stubborn problems, using software to extract business rules buried inside legacy mainframes before rebuilding them as products for a modern core platform.

The companies have integrated AWS Transform with Thought Machine’s Vault Forge tooling, creating a migration workflow intended to reverse-engineer systems containing large quantities of older code, including COBOL. Rather than translating that code directly into a newer programming language, the process is designed to identify the underlying rules governing interest, fees, repayments, account lifecycles, and other financial-product behaviour.

AWS Transform handles discovery and reverse engineering before converting identified logic into structured requirements. Those specifications are then consolidated to remove duplicated product variants, after which Vault Forge uses specialised agents running through Amazon Bedrock to generate Python products, test suites, and configuration for Thought Machine’s Vault core-banking platform.

Human review gates remain between generation and deployment. Engineering and risk teams retain explicit sign-off authority, while generated products can be placed into a Vault sandbox for further testing before any production migration takes place. Thought Machine says the approach can compress work that traditionally stretches over years, although that claim relates to parts of the migration process rather than the complete replacement of a bank’s operational core.

Legacy systems contain business history

The attraction of extracting rules rather than translating code is that old banking applications frequently contain much more than obsolete syntax. Decades of product changes, customer exceptions, regulatory adjustments, mergers, pricing decisions, and workarounds can become embedded inside applications whose original documentation is incomplete or no longer reflects production behaviour.

Replacing such a system is difficult partly because banks need to know which behaviours are deliberate and which merely survive because nobody wants to risk removing them. Translating every line can preserve unnecessary complexity, whereas redesigning the system without understanding the old logic can accidentally discard rules that remain commercially or legally important.

The Thought Machine and AWS workflow tries to place AI in that gap, using agents to reconstruct business intent before a new product is created. That is a more demanding task than generating boilerplate software because an apparently small difference in an interest calculation or repayment sequence can alter customer balances and regulatory outcomes.

Automated consolidation consequently introduces judgement as well as efficiency. Hundreds of historical product variations may genuinely be redundant, but some differences may represent customer rights, contractual terms, jurisdictional requirements, or past remediation. Banks will still need domain specialists who can determine whether apparently similar products should be merged rather than accepting an agent’s abstraction as authoritative.

Migration risk moves rather than disappears

Core banking modernisation has often been sold as a route away from expensive mainframes and tightly coupled software, although replacing technology underneath deposits, payments, loans, and customer accounts remains one of the least forgiving forms of enterprise transformation. A failed migration can affect money movement, regulatory reporting, customer access, and the accounting records on which the institution itself depends.

AI can reduce the effort involved in understanding source systems and generating target configurations, but it does not remove the other components of a migration. Customer and transaction data still have to be mapped, cleansed, reconciled, moved, and verified, while surrounding applications must be reconnected and operational teams prepared for a new platform.

Thought Machine’s Vault architecture separates financial-product logic from the underlying database infrastructure and represents products as Python code. That makes generated logic easier to test and version than rules scattered through older applications, while also creating a clearer boundary between the product definition and the platform running it.

The model potentially changes the economics of an early migration phase. Banks have traditionally spent considerable time inventorying estates, documenting behaviour, and commissioning teams to interpret source systems before major implementation begins. Automating more of that discovery could allow institutions to establish whether a migration is viable before committing to the full programme.

Thought Machine is supporting the launch with a Core Modernisation Accelerator programme in which its architects and engineers work alongside bank teams. That services component is revealing because the difficult part of core modernisation has rarely been the absence of a target software product alone; it is the organisational work required to move from one operating model to another without interrupting the institution.

AWS also gains a route for its Transform tooling into financial infrastructure, extending the use of AI modernisation agents beyond ordinary application estates. Running the pipeline inside a customer’s AWS environment is intended to keep source code and extracted business logic within the bank’s controlled cloud environment, although institutions will still need to assess model access, logging, security, auditability, and regulatory treatment.

The more interesting question is whether AI can make banks more willing to begin core replacements repeatedly postponed because nobody can fully explain the old estate. Extracting rules automatically does not make a migration safe on its own, but it could reduce the uncertainty surrounding what the existing software actually does.